Feature selection by recursive binary gravitational search algorithm optimization for cancer classification

• Published in: Soft computing (Berlin, Germany) Vol. 24; no. 6; pp. 4407 - 4425
• Main Authors: Han, Xiaohong, Li, Dengao, Liu, Ping, Wang, Li
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.03.2020; Springer Nature B.V
• Subjects: Accuracy; Artificial Intelligence; Biological properties; Cancer; Classification; Computational Intelligence; Control; Datasets; Discriminant analysis; DNA chips; Engineering; Feature selection; Gene expression; Genes; Genetic algorithms; Machine learning; Mathematical Logic and Foundations; Mechatronics; Medical diagnosis; Methodologies and Application; Methods; Optimization; Regression analysis; Robotics; Search algorithms; Support vector machines; Variance analysis; Cancer classification; Gravitational search algorithm; Gene selection
• ISSN: 1432-7643; 1433-7479
• DOI: 10.1007/s00500-019-04203-z

DNA microarray technology has become a prospective tool for cancer classification. However, DNA microarray datasets typically have very large number of genes (usually more than tens of thousands) and less number of samples (often less than one hundred). This raises the issue of getting the most relevant genes prior to cancer classification. In this paper, we have proposed a two-phase feature selection method for cancer classification. This method selects a low-dimensional set of genes to classify biological samples of binary and multi-class cancers by integrating ReliefF with recursive binary gravitational search algorithm (RBGSA). The proposed RBGSA refines the gene space from a very coarse level to a fine-grained one at each recursive step of the algorithm without degrading the accuracy. We evaluate our method by comparing it with state-of-the-art methods on 11 benchmark microarray datasets of different cancer types. Comparison results show that our method selects only a small number of genes while yielding substantial improvements in accuracy over other methods. In particular, it achieved up to 100% classification accuracy for 7 out of 11 datasets with a very small size of gene subset (up to < 1.5%) for all 11 datasets.